Blind Source Separation by Local Interaction of Output Signals

We compare independent component analysis (ICA) [Bell and Sejnowski, 1995] to an alternative method [Fisher and Principe, 1996] for blind source separation of instantaneous linear mixtures. The method and its application to blind source separation of instantaneous linear mixtures is reviewed. Empirical results separating sources of varying kurtosis and a limited number of samples are presented. We demonstrate empirically that despite the additional computational cost of the method presented, significantly better performance can be achieved for a small number of samples and when the kurtosis of the sources is sub-gaussian. The method is also of interest as it can be extended easily to any differentiable nonlinear mapping.